Project Definition¶
Project Overview¶
The project is aimed at developing an algorithm to identify the breed of a dog. Moreover, if the input image is a human one, the algorithm should rightly identify it so, and also spit out the closest resembling dog breed.
The chosen domain of the problem is image classification using Deep Learnining - specifically, using Convolutional Neural Networks.
The origin of the image classification problem using Deep Learning, in general, links back to the ImageNet Large Scale Visual Recognition Challenge. This project, specifically, was originally developed by Udacity as one of its various options for a Data Scientist Capstone Project. The goal of this exercise is to solve the problem by building a dog breed classification model.
The data sets used in the project are:
- Images of different breeds of dogs (133 of them).
- Image of human faces.
The instructions to download the images are provided in the README file of this project.
Problem Statement¶
The specifics of the problem is as follows:
- Given an input image, identify whether a dog or a human is present in the image.
- If a dog has been identified, classify it as belonging to one of the 133 different dog breeds.
- If a human has been identifed, spit out the most resembling dog breed (fun part).
Strategy for solving the problem
We will solve the problem with an increasing level of complexity:
- Step 0: Import Datasets
- Step 1: Detect Humans
- Step 2: Detect Dogs
- Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
- Step 4: Use a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 6: Build a Complete Algorithm
- Step 7: Algorithm Evaluation
The expected solution is the one that encapsulates and pieces together the various steps defined above to come up with an algorithm that would, for most of the time, rightly classify the input image into one of the 133 categories.
Evaluation Metrics¶
We use a straight-forward metric for evaluating classification problems - accuracy. We could have looked at other metrics, such as, precision and recall. But, given the high cardinality of the categories, we would instead stick to accuracy (which is simply defined as the number of correct predictions divided by the total predictions).
Analysis¶
Data Exploration & Visualization¶
Import necessary modules
import glob
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline
Extract different dog breeds available in train, valid and test data sets.
train_categories = glob.glob('dogImages/train/*')
valid_categories = glob.glob('dogImages/valid/*')
test_categories = glob.glob('dogImages/test/*')
Check for the number of dog breeds available in each of train, valid and test data sets.
len(train_categories)
len(valid_categories)
len(test_categories)
Let's plot some sample images to get a sense of the data we are dealing with.¶
from skimage import io
def sample_plot(array):
"""
Function to plot sample images
Input:
array (list): contains folders of dog breeds
Output:
plots images of random sample
"""
sample_breed = np.random.choice(array, 10)
fig, ax = plt.subplots(2, 5, figsize=(16, 8))
for en, s in enumerate(sample_breed):
sample_image = np.random.choice(glob.glob(s+'/*'))
ax[en//5][en%5].imshow(io.imread(sample_image))
ax[en//5][en%5].axis('off')
ax[en//5][en%5].set_title(s.split('/')[-1].split('.')[1])
plt.title('Random sample of images', fontsize=17)
plt.show()
Sample Images from Train Data Set¶
sample_plot(train_categories)
sample_plot(train_categories)
Sample Images from Valid Data Set¶
sample_plot(valid_categories)
sample_plot(valid_categories)
Sample Images from Test Data Set¶
sample_plot(test_categories)
sample_plot(test_categories)
Let's look at the distribution of image sizes¶
from tqdm import tqdm_notebook as tqdm
def get_images_size_distribution(array):
"""
Function to get the image size distribution
Input:
array (list): contains folders of dog breeds for which image size distribution is to be calculated
Returns:
list containing image sizes
"""
images_sizes = []
for a in tqdm(array):
for i in glob.glob(a+'/*'):
try:
images_sizes.append(io.imread(i).shape)
except:
pass
return images_sizes
train_image_sizes = get_images_size_distribution(train_categories)
valid_image_sizes = get_images_size_distribution(valid_categories)
test_image_sizes = get_images_size_distribution(test_categories)
def plot_image_size_distribution(array):
"""
Function to plot distribution of image features
Input:
array (list): contains image sizes
Outputs:
plot showing the distribution of image features
"""
fig, ax = plt.subplots(2, 1, figsize=(16, 12))
for i in range(2):
dim_distribution = []
for a in array:
try:
dim_distribution.append(a[i])
except:
pass
ax[i].hist(dim_distribution, bins=100)
ax[i].set_title('Data distribution along dimension {}'.format(str(i)))
ax[i].set_xlabel('Pixels along dimension {}'.format(i))
ax[i].set_ylabel('Frequency')
plt.show()
Plot of Image Size Distribution in Train Data Set¶
plot_image_size_distribution(train_image_sizes)
Plot of Image Size Distribution in Valid Data Set¶
plot_image_size_distribution(valid_image_sizes)
Plot of Image Size Distribution in Test Data Set¶
plot_image_size_distribution(test_image_sizes)
We see that roughly in all the three data sets - train, valid and test - the height and width of the images are centered around ~500 pixels.
Let's plot all the distributions in one plot for easy comparision¶
def plot_image_size_distribution_complete(array1, array2, array3):
"""
Function to plot distribution of image features
Input:
array1 (list): contains image sizes of training data set
array2 (list): contains image sizes of validation data set
array3 (list): contains image sizes of testing data set
Outputs:
plot showing the distribution of image features
"""
fig, ax = plt.subplots(2, 1, figsize=(16, 12))
labels = ['train', 'valid', 'test']
for i in range(2):
dim_distribution = []
for en, array in enumerate([array1, array2, array3]):
for a in array:
try:
dim_distribution.append(a[i])
except:
pass
ax[i].hist(dim_distribution, bins=100, alpha=0.3, label=labels[en], normed=True)
ax[i].set_title('Data distribution along dimension {}'.format(str(i)))
ax[i].set_xlabel('Pixels along dimension {}'.format(i))
ax[i].set_ylabel('Frequency in %')
ax[i].legend()
plt.show()
plot_image_size_distribution_complete(train_image_sizes, valid_image_sizes, test_image_sizes)
We see that the distribution of height and width of the images for all the three data sets are roughly equal
def plot_image_pixel_distribution_complete(array1, array2, array3):
"""
Function to plot distribution of image pixels
Input:
array1 (list): contains image sizes of training data set
array2 (list): contains image sizes of validation data set
array3 (list): contains image sizes of testing data set
Outputs:
plot showing the distribution of image pixels
"""
fig, ax = plt.subplots(1, 1, figsize=(16, 8))
labels = ['train', 'valid', 'test']
pixel_distribution = []
for en, array in enumerate([array1, array2, array3]):
for a in array:
try:
pixel_distribution.append(a[0]*a[1]*a[2])
except:
pass
ax.ticklabel_format(useOffset=False, style='plain')
ax.hist(pixel_distribution, bins=300, alpha=0.3, label=labels[en], normed=True)
ax.set_title('Pixel Distribution Median: {}'.format(int(np.median(pixel_distribution))), fontsize=17)
ax.set_xlabel('# Pixels')
ax.set_ylabel('Frequency in %')
ax.legend()
plt.show()
return pixel_distribution
test = plot_image_pixel_distribution_complete(train_image_sizes, valid_image_sizes, test_image_sizes)
We see that the median pixels for all the three data sets - train, valid and test - is centered around ~700K pixels. This also confirms that we are dealing with color images because from the above graphs we see that the median values for height and width of the images are 500 each. The depth of the color images is 3. Therefore, 500 X 500 X 3 = 750K, which roughly equals the median value of pixel distribution
Check for data distribution across dog breeds for each of train, valid and test data sets.|
def get_data_distribution(array):
"""
Function to extract data distribution across dog breeds
Input:
array (list): glob file paths for train/valid/test data sets
Output:
dog_breeds (list)
nimages (list)
"""
dog_breeds, nimages = [], []
for a in array:
dog_breeds.append(a.split('/')[-1].split('.')[1])
nimages.append(len(glob.glob(a+'/*')))
return dog_breeds, nimages
dog_breeds_train, nimages_train = get_data_distribution(train_categories)
dog_breeds_valid, nimages_valid = get_data_distribution(valid_categories)
dog_breeds_test, nimages_test = get_data_distribution(test_categories)
Ensure that the same dog breeds are available across train, valid and test data sets
len(set(dog_breeds_train).intersection(set(dog_breeds_valid)).intersection(set(dog_breeds_test)))
Let's look at the data distribution across train, valid and test data sets
fig, ax = plt.subplots(figsize=(16, 8))
ax.bar(['train', 'valid', 'test'], [sum(nimages_train), sum(nimages_valid), sum(nimages_test)],
color=['r', 'b', 'g'], alpha=0.6)
plt.title('Data Distribution (Absolute)', fontsize=15)
ax.set_ylabel('No. images')
plt.show()
nimages_total = sum(nimages_train) + sum(nimages_valid) + sum(nimages_test)
fig, ax = plt.subplots(figsize=(16, 8))
ax.bar(np.array(['train', 'valid', 'test']), np.array([sum(nimages_train), sum(nimages_valid), sum(nimages_test)])*100/nimages_total,
color=['r', 'b', 'g'], alpha=0.6)
plt.title('Data Distribution (Relative)', fontsize=15)
ax.set_ylabel('% of images')
plt.show()
There is, approximately, a 80%-10%-10% split across training, validation and testing datasets
Now, let's look at the distribution across categories.
nimages_total = sum(nimages_train) + sum(nimages_valid) + sum(nimages_test)
fig, ax = plt.subplots(figsize=(20, 10))
ax.bar(np.array(dog_breeds_train), np.array(nimages_train)*100/sum(nimages_train), alpha=0.6)
plt.title('Data Distribution (Train)', fontsize=15)
ax.set_ylabel('% of images')
ax.xaxis.set_tick_params(rotation=90)
plt.show()
nimages_total = sum(nimages_train) + sum(nimages_valid) + sum(nimages_test)
fig, ax = plt.subplots(figsize=(20, 10))
ax.bar(np.array(dog_breeds_valid), np.array(nimages_valid)*100/sum(nimages_valid), alpha=0.6)
plt.title('Data Distribution (Valid)', fontsize=15)
ax.set_ylabel('% of images')
ax.xaxis.set_tick_params(rotation=90)
plt.show()
nimages_total = sum(nimages_train) + sum(nimages_valid) + sum(nimages_test)
fig, ax = plt.subplots(figsize=(20, 10))
ax.bar(np.array(dog_breeds_test), np.array(nimages_test)*100/sum(nimages_test), alpha=0.6)
plt.title('Data Distribution (Test)', fontsize=15)
ax.set_ylabel('% of images')
ax.xaxis.set_tick_params(rotation=90)
plt.show()
The distribution across categories for each of train, valid and test data sets also rougly follow the same pattern. Also, although there seems to be a variance in the data distribution of different categories (0.5% to 1.2%), it is not exceptionally worrying.
Now, let's look at the number of human faces in our data set
len(glob.glob('lfw/*/*'))
This data set is not split across train, valid and test sets.
Onto modeling!
Methodology¶
Please note: Relevant and necessary pre-processing steps required are mentioned as and when necessary.
Step 0: Import Datasets¶
Import Dog Dataset¶
In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:
train_files,valid_files,test_files- numpy arrays containing file paths to imagestrain_targets,valid_targets,test_targets- numpy arrays containing onehot-encoded classification labelsdog_names- list of string-valued dog breed names for translating labels
from sklearn.datasets import load_files
from keras.utils import np_utils
import numpy as np
from glob import glob
# define function to load train, test, and validation datasets
def load_dataset(path):
data = load_files(path)
dog_files = np.array(data['filenames'])
dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
return dog_files, dog_targets
# load train, test, and validation datasets
train_files, train_targets = load_dataset('dogImages/train')
valid_files, valid_targets = load_dataset('dogImages/valid')
test_files, test_targets = load_dataset('dogImages/test')
# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]
# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
Import Human Dataset¶
In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.
import random
random.seed(8675309)
# load filenames in shuffled human dataset
human_files = np.array(glob("lfw/*/*"))
random.shuffle(human_files)
# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
Step 1: Detect Humans¶
We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.
In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
import cv2
import matplotlib.pyplot as plt
%matplotlib inline
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
img = cv2.imread(human_files[3])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
Human Face Detector¶
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
Assessmet of the Human Face Detector¶
Let's look at the following questions:
- What percentage of the first 100 images in
human_fileshave a detected human face? - What percentage of the first 100 images in
dog_fileshave a detected human face?
Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer:
- Out of the first 100 images from
human_files,face_detectoridentified all 100 images as actually containing human faces in it. - Out of the first 100 images from
dog_files,face_detectoridentified 12 images as containing human faces (instead of dog faces) in it.
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
humans_detector=[face_detector(h) for h in human_files_short]
dogs_detector=[face_detector(d) for d in dog_files_short]
print('Percentage of human faces detected in human_files_short - {}%'.format(sum(humans_detector)))
print('Percentage of human faces detected in dog_files_short - {}%'.format(sum(dogs_detector)))
Food for thought: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). Is this a reasonable expectation to pose on the user?
The expectation of users providing a clear view of thier faces isn't a reasonable expectation to pose onto the users. Especially, in today's world where cameras are ubiquitous, users click photos from various angles and poses. A good algorithm is expected to catch-in all these images, and not just a select few where the images provide a clear view of faces. I think the use of convoultional neural nets should solve this problem since it has in-built translational invariance. I'm excited to see if convNet solves this!
Step 2: Detect Dogs¶
In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.
from keras.applications.resnet50 import ResNet50
# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')
Pre-process the Data¶
When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape
$$ (\text{nb_samples}, \text{rows}, \text{columns}, \text{channels}), $$
where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.
The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape
$$ (1, 224, 224, 3). $$
The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape
$$ (\text{nb_samples}, 224, 224, 3). $$
Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in our dataset!
from keras.preprocessing import image
from tqdm import tqdm
def path_to_tensor(img_path):
# loads RGB image as PIL.Image.Image type
img = image.load_img(img_path, target_size=(224, 224))
# convert PIL.Image.Image type to 3D tensor with shape (224, 224, 3)
x = image.img_to_array(img)
# convert 3D tensor to 4D tensor with shape (1, 224, 224, 3) and return 4D tensor
return np.expand_dims(x, axis=0)
def paths_to_tensor(img_paths):
list_of_tensors = [path_to_tensor(img_path) for img_path in tqdm(img_paths)]
return np.vstack(list_of_tensors)
Making Predictions with ResNet-50¶
Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. For more information, please check the code for preprocess_input here.
Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.
By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.
from keras.applications.resnet50 import preprocess_input, decode_predictions
def ResNet50_predict_labels(img_path):
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
Dog Detector¶
While looking at the dictionary, we will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).
We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
prediction = ResNet50_predict_labels(img_path)
return ((prediction <= 268) & (prediction >= 151))
Assessment of the Dog Detector¶
Let's look at the following questions:
- What percentage of the images in
human_files_shorthave a detected dog? - What percentage of the images in
dog_files_shorthave a detected dog?
Answer:
- 0 percentage of the images in
human_files_shorthas a dog detected in it. - 100% of the images in
dog_files_shorthave dogs detected in them.
### TODO: Test the performance of the dog_detector function
### on the images in human_files_short and dog_files_short.
human_detector_resnet = [dog_detector(h) for h in human_files_short]
dog_detector_resnet = [dog_detector(d) for d in dog_files_short]
print('Percentage of dogs detected in human_files_short - {}%'.format(sum(human_detector_resnet)))
print('Percentage of dogs detected in dog_files_short - {}%'.format(sum(dog_detector_resnet)))
Step 3: Create a CNN to Classify Dog Breeds (from Scratch)¶
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, we will create a CNN that classifies dog breeds. Our goal is to attain a test accuracy of at least 1%. In Step 5 of this notebook, we will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
We are careful with addingg too many trainable layers! More parameters means longer training, which means we are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; we can extrapolate this estimate to figure out how long it will take for our algorithm to train.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
 |
 |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
 |
 |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
 |
 |
 |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Pre-process the Data¶
We rescale the images by dividing every pixel in every image by 255.
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
Model Architecture¶
A Sample Model Architecture
I have taken the basic guidelines from the sample architecture displayed above. However, in the first convolution layer, I've increased the filter size to (3,3). This is because, in the initial stages of the architecture, covering a larger filter area seemed more intuitive to me as doing so will give us more freedom in capturing patterns occuring over larger area. Gradually increasing the depth by using more filters in the subsequent layers, and gradually decreasing the height and width of the image by using max-pooling layers, as suggested by the sample architecture, seemed sensible to me. Increasing depth will capture more patterns, and decreasing height and width will reduce the number of parameters that we will have to fiddle with. Rectified linear units are used for activation functions as they allow the gradient information to flow smoothly during backpropagation. The softmax activation at the final layer is basically used to spit out probabilities. The number of additional parameters per each layer is taken into consideration while deciding on the number of layers.
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
model = Sequential()
model.add(Conv2D(filters=16, kernel_size=(3,3), padding='valid', activation='relu',
input_shape=(224, 224, 3)))
model.add(MaxPooling2D((2,2)))
model.add(Conv2D(filters=32, kernel_size=(2,2), padding='valid', activation='relu'))
model.add(MaxPooling2D(2,2))
model.add(Conv2D(filters=64, kernel_size=(2,2), padding='valid', activation='relu'))
model.add(Flatten())
model.add(Dense(133, activation='softmax'))
### TODO: Define your architecture.
model.summary()
Compile the Model¶
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
Train the Model¶
from keras.callbacks import ModelCheckpoint
### specify the number of epochs that you would like to use to train the model.
epochs = 5
### Do NOT modify the code below this line.
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.from_scratch.hdf5',
verbose=1, save_best_only=True)
model.fit(train_tensors, train_targets,
validation_data=(valid_tensors, valid_targets),
epochs=epochs, batch_size=20, callbacks=[checkpointer], verbose=1)
Load the Model with the Best Validation Loss¶
model.load_weights('saved_models/weights.best.from_scratch.hdf5')
Test the Model¶
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0))) for tensor in test_tensors]
# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']
Model Architecture¶
The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.
VGG16_model = Sequential()
VGG16_model.add(GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(Dense(133, activation='softmax'))
VGG16_model.summary()
Compile the Model¶
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
Train the Model¶
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5',
verbose=1, save_best_only=True)
VGG16_model.fit(train_VGG16, train_targets,
validation_data=(valid_VGG16, valid_targets),
epochs=20, batch_size=20, callbacks=[checkpointer], verbose=1)
Load the Model with the Best Validation Loss¶
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')
Test the Model¶
Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0))) for feature in test_VGG16]
# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
from extract_bottleneck_features import *
def VGG16_predict_breed(img_path):
# extract bottleneck features
bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = VGG16_model.predict(bottleneck_feature)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)¶
We will now use transfer learning to create a CNN that can identify dog breed from images.
In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, we will use the bottleneck features from a different pre-trained model. To make things easier, we have pre-computed the features for all of the networks that are currently available in Keras:
- VGG-19 bottleneck features
- ResNet-50 bottleneck features
- Inception bottleneck features
- Xception bottleneck features
The files are encoded as such:
Dog{network}Data.npz
where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.
Obtain Bottleneck Features¶
In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:
bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']ResNet-50¶
### Obtain bottleneck features from another pre-trained CNN.
bottleneck_features = np.load('bottleneck_features/DogResnet50Data.npz')
train_resnet = bottleneck_features['train']
valid_resnet = bottleneck_features['valid']
test_resnet = bottleneck_features['test']
Model Architecture¶
Answer:
I basically wanted to test all the four architectures - VGG19, ResNet, Inception and Xception - keeping all the other parameters constant. Since we've already tested VGG16 with rms_prop optimizer, categorical_crossentropy loss function, 20 epochs, batch size of 20, an additional GlobalAveragePooling layer and a softmax dense layer at the end, I've used the same parameters for different bottleneck features. The reason why I think the architechtures are suitable for the current problem is because all these architectures' weights were fine-tuned usind ImageNet data which contained all the different breeds of dogs. Hence, these architectures have already learnt to differentiate amongst the various dog breeds (and more). As is apparent, the Xception arcitecture performs the best on the test set.
### Define architecture.
resnet = Sequential()
resnet.add(GlobalAveragePooling2D(input_shape=train_resnet.shape[1:]))
resnet.add(Dense(133, activation='softmax'))
resnet.summary()
Compile the Model¶
resnet.compile(optimizer='rmsprop', metrics=['accuracy'], loss='categorical_crossentropy')
Train the Model¶
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.resnet.hdf5',
verbose=1, save_best_only=True)
resnet.fit(train_resnet, train_targets,
validation_data=(valid_resnet, valid_targets),
verbose=1, epochs=20, batch_size=20,callbacks=[checkpointer])
Load the Model with the Best Validation Loss¶
resnet.load_weights('saved_models/weights.best.resnet.hdf5')
Test the Model¶
### Calculate classification accuracy on the test dataset.
resnet_predictions = [np.argmax(resnet.predict(np.expand_dims(tensor,axis=0))) for tensor in test_resnet]
resnet_accuracy = 100*np.sum(np.array(resnet_predictions)==np.argmax(test_targets, axis=1))/len(test_targets)
print("(Resnet) Test accuracy: %.4f%%" % resnet_accuracy)
VGG19¶
bottleneck_features = np.load('bottleneck_features/DogVGG19Data.npz')
train_VGG19 = bottleneck_features['train']
valid_VGG19 = bottleneck_features['valid']
test_VGG19 = bottleneck_features['test']
VGG19 = Sequential()
VGG19.add(GlobalAveragePooling2D(input_shape=train_VGG19.shape[1:]))
VGG19.add(Dense(133, activation='softmax'))
VGG19.summary()
VGG19.compile(optimizer='rmsprop', metrics=['accuracy'], loss='categorical_crossentropy')
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG19.hdf5', verbose=1, save_best_only=True)
VGG19.fit(train_VGG19, train_targets, validation_data=(valid_VGG19, valid_targets), epochs=20, batch_size=20,
verbose=1, callbacks=[checkpointer])
VGG19.load_weights('saved_models/weights.best.VGG19.hdf5')
VGG19_predictions = [np.argmax(VGG19.predict(np.expand_dims(tensor, axis=0))) for tensor in test_VGG19]
VGG19_accuracy = 100*np.sum(np.array(VGG19_predictions)==np.argmax(test_targets, axis=1))/len(test_targets)
print(' (VGG19) Test accuracy: %.4f%%' % VGG19_accuracy)
Inception¶
bottleneck_features = np.load('bottleneck_features/DogInceptionV3Data.npz')
train_inception = bottleneck_features['train']
valid_inception = bottleneck_features['valid']
test_inception = bottleneck_features['test']
inception = Sequential()
inception.add(GlobalAveragePooling2D(input_shape=train_inception.shape[1:]))
inception.add(Dense(133, activation='softmax'))
inception.summary()
inception.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.inception.hdf5', verbose=1, save_best_only=True)
inception.fit(train_inception, train_targets, validation_data=(valid_inception, valid_targets),
epochs=20, batch_size=20, verbose=1, callbacks=[checkpointer])
inception.load_weights('saved_models/weights.best.inception.hdf5')
inception_predictions = [ np.argmax(inception.predict(np.expand_dims(tensor, axis=0))) for tensor in test_inception]
inception_accuracy = 100*np.sum(np.array(inception_predictions)==np.argmax(test_targets, axis=1))/len(test_targets)
print(' (Inception) Test accuracy: %.4f%%' % inception_accuracy)
Xception¶
bottleneck_features = np.load('bottleneck_features/DogXceptionData.npz')
train_xception = bottleneck_features['train']
valid_xception = bottleneck_features['valid']
test_xception = bottleneck_features['test']
xception = Sequential()
xception.add(GlobalAveragePooling2D(input_shape=train_xception.shape[1:]))
xception.add(Dense(133, activation='softmax'))
xception.summary()
xception.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.xception.hdf5', save_best_only=True, verbose=1)
xception.fit(train_xception, train_targets, validation_data=(valid_xception, valid_targets), epochs=20, batch_size=20,
callbacks=[checkpointer], verbose=1)
xception.load_weights('saved_models/weights.best.xception.hdf5')
xception_predictions = [ np.argmax(xception.predict(np.expand_dims(tensor, axis=0))) for tensor in test_xception]
xception_accuracy = 100*np.sum(np.array(xception_predictions==np.argmax(test_targets, axis=1)))/len(test_targets)
print(' (Xception) Test accuracy: %.4f%%' % xception_accuracy)
Predict Dog Breed with the Model¶
Let's now write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.
Similar to the analogous function in Step 5, our function should have three steps:
- Extract the bottleneck features corresponding to the chosen CNN model.
- Supply the bottleneck features as input to the model to return the predicted vector. Note that the argmax of this prediction vector gives the index of the predicted dog breed.
- Use the
dog_namesarray defined in Step 0 of this notebook to return the corresponding breed.
The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to our chosen CNN architecture, you need to use the function
extract_{network}
where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.
#Note: I'm not re-building the model as I already have the model built.
def predict_dog_breed(image_path, model=xception, dog_names=dog_names):
"""
Fuction to predict the dog breed
Inputs:
image_path: path to the input image
model: model to be used for prediction
dog_names: array containing dog breeds
Returns:
predicted dog breed
"""
bottleneck_features=extract_Xception(path_to_tensor(image_path))
prediction=dog_names[np.argmax(model.predict(bottleneck_features),axis=1)[0]]
return prediction
predict_dog_breed(dog_files_short[0])
Indeed, the prediction is correct.
dog_names[np.argmax(train_targets[0])]
Step 6: Build a Complete Algorithm¶
Finally, we build an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
- if a dog is detected in the image, return the predicted breed.
- if a human is detected in the image, return the resembling dog breed.
- if neither is detected in the image, provide output that indicates an error.
def identify_image(image_path, cnn=xception):
"""
Function to predict the dog breed of an input image
Inputs:
image_path: path to the image on which prediction is to be made
cnn: model to be used for prediction
Output:
Model prediction
"""
#extract bottleneck features
bottleneck_features=extract_Xception(path_to_tensor(image_path))
#predict the dog breed
predictions=cnn.predict(bottleneck_features)
confidence=predictions.max()
dog_breed=predict_dog_breed(image_path)
img=cv2.imread(image_path)
img_rgb=cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
#output the prediction
if dog_detector(image_path):
print('Hello, dog!')
plt.imshow(img_rgb)
print('You belong to the breed '+dog_breed+' with '+str(round(confidence*100,2))+'% confidence')
elif face_detector(image_path):
print('Hello, human!')
plt.imshow(img_rgb)
print('You look like a '+dog_breed+' with '+str(round(confidence*100,2))+'% confidence')
else:
print('Error! Neither dogs nor humans detected in the image.')
plt.imshow(img_rgb)
plt.show()
identify_image('Image uploaded from iOS.jpg')
Results¶
Model Evaluation and Validation¶
Before we go ahead and evaluate the algorithm, let's summarize the accuracy metric for the various architectures employed.
| Architecture | Trainable Parameters | Train Accuracy | Valid Accuracy | Test Accuracy |
|---|---|---|---|---|
| Custom | 24,831,909 | 97.6% | 5.03% | 4.19% |
| VGG16 | 68,229 | 63.58% | 49.7% | 49.76% |
| VGG19 | 68,229 | 61.74% | 48.14% | 50.0% |
| ResNet-50 | 272,517 | 99.82% | 82.75% | 79.9% |
| Inception | 272,517 | 98.88% | 86.23% | 77.75% |
| Xception | 272,517 | 97.98% | 84.91% | 83.25% |
identify_image('images/german_shepard.jpeg')
identify_image('images/pitbull.jpg')
identify_image('images/bull_dog.jpeg')
identify_image('images/cat.jpg')
identify_image('images/human.jpg')
identify_image('images/human1.jpg')
Justification¶
The model's output just fell a tad short of our expectations. The model correctly predicts 5 out of the 6 input images. It correctly identified the dog breeds of German shepherd and bull dog. Also, it was right in idenfying the human faces and cats in the input images. However, it got confused with pitbull image, and actually thinks that it belongs to a type of French breeds - Dogue_de_bordeaux. But, the good part is that the confidence in this prediction is low.
As was thoroughly discussed above, the transfer learning approach made it possible for us to achive this level of accuracy over our initial custom CNN model built from scratch.
Conclusion¶
Reflection¶
In summary, we've built a function to identify_image which does the following tasks:
- Check for the presence of a dog using
dog_detectorfunction, which in turn usesResNet50architecture to check if the prediction falls between 151-268 (the key ranges for dog breeds). If so, then the function outputs the dog breed predicted using thepredict_dog_breedfunction, which in turn uses a variant of theXceptionarchitecture built using transfer learning. - If a dog is not detected, we check for the presence of a human in the input image using
face_detectorfunction. Theface_detectorfunction uses OpenCV's Haar feature-based cascade classifiers to detect the presence of a human face in the input image. If a human face is found, we also output the closest resembling dog breed using the samepredict_dog_breedfunction described above. - If neither a human nor a dog is found in the image, we output the error message.
The aspects of the projects that I particularly found interesting are:
- The efficacy of transfer learning: how by using bottleneck features and building minimal layers on top of them we could achieve an impressive boost up in the accuracy.
- The ability to chain independent modules to achieve an impressive application. For this specific project, we've used OpenCV's face detector algorithm and clubbed it with Deep Learning architectures to come up with an interesting and fun application.
Future Work for Improvement¶
The three potential points of improvement in the algorithm are:
- The algorithm can be improved to be able to differentiate between fully grown dogs and puppies. This would be an additional imporovement to our present model wherein we not only classify the dog breeds, but also whether the dog is a fully grown one or a puppy.
- It can also be fine-tuned to correctly identify different breeds of dogs/cats in the same image. Incorporating the will allow the model to identify multiple dogs and their respective breeds in the input image and not just a single one.
- Also, for human input images, the model can be further improved to bring the confidence down to 0%. Although the present model rightly classifies human images with ~10-15% confidence, striving to bring this further down will help improve the model performance.